Skip to content

Run the [embed] e2e scripts in CI to catch ORT/onnxruntime_go coupling breaks #12

Description

@johnsabath

The problem

The e2e suite already covers the real embedding path with txtar scripts (cmd/semantic/testdata/search-quality.txtar, gated on a [embed] condition), but by design CI never runs them — TestE2E's own comment says so, to keep CI offline and avoid a ~127 MB download on every run. That means a break in the onnxruntime_go/OrtVersion coupling — exactly what #11 fixed — has nothing to catch it before a release ships. It only surfaced because someone happened to run semantic search for real after tagging v0.2.0.

What you propose

A CI job that isn't part of the normal per-PR run — scheduled, or gated to release prep — that runs semantic init to fetch the real model + ONNX Runtime, then go test ./cmd/semantic/... -run TestE2E so the existing [embed] txtar scripts actually execute. No new test mechanism needed, just somewhere for the one that exists to run against a real model before a tag goes out.

What else you considered

A cheaper smoke check that only loads the ONNX Runtime shared library (skips the 127 MB model download) would catch an API-version mismatch specifically, but not the coupling breaks the txtar scripts are built to catch generally.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions